Kubecost y OpenCost: FinOps nativo en Kubernetes

javi
Gráfico financiero ascendente con barras coloridas sobre tableta digital

OpenCost (CNCF sandbox) y Kubecost (su versión comercial) resuelven el problema básico de Kubernetes: “¿cuánto cuesta cada namespace?”. Cloud bills agrupan por recurso/tag, no por workload K8s. Kubecost mapea AWS/GCP/Azure billing a constructs K8s: pods, deployments, namespaces, labels.

Qué ofrecen

  • Cost allocation: per pod, namespace, label, deployment.
  • Optimization recommendations: right-sizing, spot, abandoned workloads.
  • Cloud billing integration: AWS CUR, GCP Billing Export, Azure.
  • Showback/Chargeback: invoices internas per team.
  • Forecasting: predicciones de coste.
  • Alerting: budget exceeded notifications.

OpenCost vs Kubecost

Aspect OpenCost Kubecost
License Apache 2.0 Free tier + paid
Cost allocation Sí Sí
Cloud billing Básico Avanzado
Multi-cluster Limitado Sí
UI Básica Rica
Support Community Commercial

OpenCost = free essentials. Kubecost = paid features (enterprise).

Instalación

Helm:

helm repo add opencost https://opencost.github.io/opencost-helm-chart
helm install opencost opencost/opencost

# Or Kubecost
helm repo add kubecost https://kubecost.github.io/cost-analyzer/
helm install kubecost kubecost/cost-analyzer \
  --set kubecostToken="your-token"

Deploy requires:

  • Prometheus scraping.
  • Cloud IAM permissions para billing APIs.

Cost allocation accuracy

Factores:

  • Prometheus data resolution: 5-minute intervals decent.
  • CSV / billing integration: más preciso.
  • Historical data: 15-30 días típicamente.
  • Idle cost distribution: allocation de nodos subutilizados.

Accuracy típica: 85-95% vs real invoice.

Recommendations

Kubecost sugiere:

  • Rightsizing: reduce requests (pods overprovisioned).
  • Spot instances: workloads tolerant.
  • Reserved instances: long-running steady.
  • Abandoned workloads: deployments 0 traffic.
  • Unhealthy pods: consuming sin delivering value.

Tipical savings identified: 20-40% primer month.

Showback interno

Report a teams:

Team Frontend
  - namespace: frontend-prod: $1,230/mes
  - namespace: frontend-staging: $180/mes
  - total: $1,410

Team Backend
  - namespace: api-prod: $3,450/mes
  - namespace: workers: $890/mes
  - total: $4,340

Transparency accountability incentiva efficiency.

Cloud integration

AWS

  • Cost and Usage Report (CUR) en S3.
  • IAM role con ce:GetCostAndUsage.
  • RDS, EKS, ELB: pricing data.

GCP

  • Billing Export a BigQuery.
  • Service account con billing.viewer.

Azure

  • Cost Management Export.
  • Service principal con billing reader.

Setup toma unas horas por cloud.

Métricas específicas Kubernetes

Kubecost tracks:

  • CPU allocation vs usage por pod.
  • Memory allocation vs usage.
  • Storage PVC por namespace.
  • Network egress (limited visibility).
  • GPU utilization.

Alerting

alerts:
  - name: namespace-over-budget
    type: budget
    namespace: production
    threshold: 5000  # $5k/month
    window: 30d
    recipients:
      - slack:devops
      - email:finance@company.com

Proactive cost alerts.

Export data

Kubecost expone metrics:

  • Prometheus endpoint.
  • API REST para cost data.
  • CSV export.
  • Grafana dashboards oficial.

Integra con FinOps tools enterprise (Apptio, Cloudability).

Optimization workflow

  1. Baseline: medir coste actual mensual.
  2. Identify top spenders: pareto — 20% de workloads son 80% del cost.
  3. Right-size: CPU/memory requests optimization.
  4. Policy: LimitRange, ResourceQuota.
  5. Spot strategy: workloads tolerant → spot.
  6. Monitor savings: trend continuo.

Proceso iterativo cada mes-trimestre.

FinOps practices

FinOps framework phases:

  • Inform: visibility (Kubecost da esto).
  • Optimize: acting on data.
  • Operate: continuous improvement.

Kubecost herramienta. Process + culture es FinOps real.

Casos reales

  • Adobe: Kubecost enterprise-wide.
  • Ford: K8s FinOps via Kubecost.
  • Many startups: free tier sufficient.

Teams reportan 25-40% reducción de K8s spend.

Limitations

  • Network cost: ingress/egress attribution hard.
  • Shared services: e.g. Redis shared, cómo allocate.
  • Accuracy gaps con spot/reservations.
  • UI overhead: learning curve.

Opciones complementarias

  • CloudHealth, Cloudability: enterprise FinOps multi-cloud.
  • Vantage: modern SaaS.
  • Cast.ai: K8s optimization automated.

Para enterprise complex, combo Kubecost + enterprise FinOps.

Conclusión

Kubecost/OpenCost transforma K8s spend desde black box a managed cost. Para clusters >$5k/mes, ROI claro — savings suelen exceder software cost. OpenCost gratis cubre essentials; Kubecost Business+ vale la pena para features multi-cluster y cloud billing avanzado. Implementación es día de setup + semanas iterando recommendations. Para FinOps serio en K8s, tool fundamental.

Síguenos en jacar.es para más sobre FinOps, Kubernetes y cost optimization.

Entradas relacionadas